论文标题
通过神经网络引力波插值
Gravitational-wave signal-to-noise interpolation via neural networks
论文作者
论文摘要
计算信噪比（SNR）是重力波数据分析中最常见的任务之一。虽然单个SNR评估通常很快，但为整个合并事件的计算SNR可能很耗时。我们使用选定的波形模型和检测器噪声曲线计算（检测器框架）总质量，质量比和自旋大小的（检测器）总质量，质量比和自旋大小的函数来计算SNR，然后我们在此四维参数空间中使用简单的神经网络（一个多层神经perceptron）插值SNR。训练有素的网络可以在一分钟内在4核CPU上评估$ 10^6 $ snrs，中位数错误低于$ 10^{ - 3} $。这对应于$ [120，7.5 \ times10^4] $范围内的因素的平均加速度，具体取决于基础波形模型。我们训练有素的网络（和源代码）可在https://github.com/kazewong/neuralsnr上公开获取，并且很容易适应类似的多维插值问题。
Computing signal-to-noise ratios (SNRs) is one of the most common tasks in gravitational-wave data analysis. While a single SNR evaluation is generally fast, computing SNRs for an entire population of merger events could be time consuming. We compute SNRs for aligned-spin binary black-hole mergers as a function of the (detector-frame) total mass, mass ratio and spin magnitudes using selected waveform models and detector noise curves, then we interpolate the SNRs in this four-dimensional parameter space with a simple neural network (a multilayer perceptron). The trained network can evaluate $10^6$ SNRs on a 4-core CPU within a minute with a median fractional error below $10^{-3}$. This corresponds to average speed-ups by factors in the range $[120,\,7.5\times10^4]$, depending on the underlying waveform model. Our trained network (and source code) is publicly available at https://github.com/kazewong/NeuralSNR, and it can be easily adapted to similar multidimensional interpolation problems.
